Re: Heap allocation in the RTS

Are you saying allocateMightFail ignores the usual nursery size?

Right, so normally in Cmm you'd do something like `if (Hp + size > HpLim) { trigger GC }`, but allocateMightFail adds more blocks to the nursery instead. Maybe just look at the code, it's quite simple. I don't know how to check Hp in the RTS and trigger a GC. I'd do that part in Cmm as there are lots of Cmm functions that do this already (in PrimOps.cmm and maybe elsewhere). Ömer David Feuer , 20 Eyl 2018 Per, 12:50 tarihinde şunu yazdı:

I'm not sure I understand. Are you saying allocateMightFail ignores the usual nursery size? That's not my intention. It would actually be just fine to simply fail if GC would be required--I can then back off, fail out to CMM, trigger a GC there, and retry. Or I could perform an extra heap check before I start; that's a little silly, but I doubt it'll be expensive enough to really matter here.

On Thu, Sep 20, 2018, 5:42 AM Ömer Sinan Ağacan wrote:

...

allocateMightFail allocates new nursery blocks as long as you don't hit the heap limit, so it returns NULL less often than you might think. In particular, it doesn't return NULL when the nursery is full, instead it allocates a new block and adds it to the nursery.

I'd do the GC triggering part in Cmm code instead of C code -- I'm not sure if it's possible to do this in C code. There should be some functions in PrimOps.cmm that do heap allocation, maybe look there. I'd look for uses of ALLOC_PRIM. The file HeapStackCheck.cmm may also be helpful (may at least give an idea of how a GC is triggered).

Ömer

David Feuer , 20 Eyl 2018 Per, 12:34 tarihinde şunu yazdı:

...

If it returns NULL, then I need to back off what I'm doing and trigger a GC. How do I do the latter?

On Thu, Sep 20, 2018, 5:31 AM Ömer Sinan Ağacan wrote:

...

allocateMightFail does the heap check for you and returns NULL. For the current capability you can use MyCapability() in Cmm and pass the value to the RTS function you're implementing.

Ömer

David Feuer , 20 Eyl 2018 Per, 12:26 tarihinde şunu yazdı:

...

Aha! Okay. How do I get the current capability to pass to that? I guess I should probably perform a heap check before calling lookupSta bleName for simplicity, at least to start.

On Thu, Sep 20, 2018, 5:16 AM Ömer Sinan Ağacan wrote:

...

Have you seen Storage.c:allocateMightFail ?

Ömer

David Feuer , 20 Eyl 2018 Per, 11:32 tarihinde şunu yazdı: > > I'm working on re-implementing the stable name system. For the new design, it seems much cleaner to allocate stable names in lookupStableName (in rts/StableName.c) rather than in the C-- code that calls it. But I haven't seen RTS code that does heap allocation. Is it possible? If so, how do I do it? > _______________________________________________ > ghc-devs mailing list > ghc-devs@haskell.org > http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs